class: center, middle, inverse, title-slide .title[ # Analyzing gender inequality <br> in digital traces <br> and the technologies they power ] .author[ ### David Garcia <br> <em>University of Konstanz</em> ] .date[ ### slides at dgarcia.eu/ETH-Gender-2024 ] --- layout: true <div class="my-footer"><span>David Garcia - Analyzing gender inequality in digital traces and the technologies they power</span></div> --- background-image: url(figures/AboutUS.svg) background-size: 98% --- layout: true <div class="my-footer"><span><a href="https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-016-0066-4"> Women Through the Glass Ceiling: Gender Asymmetries in Wikipedia. Claudia Wagner, Eduardo Graells-Garrido, David Garcia, Filippo Menczer. EPJ Data Science (2016) </a></span></div> --- # Gender bias in digital traces: Wikipedia .pull-left[ - Fraction of male editors in Wikipedia: 91% (2011) - 5th most visited site in number of visitors (2023) - **Fundational dataset for AI** - **"The Pile" contains 20GB of Wikipedia text** ] .pull-right[ .center[] ] <div style="font-size:22pt"><span> <a href="https://ojs.aaai.org/index.php/ICWSM/article/view/14628/14477"> It's a Man's Wikipedia? Assessing Gender Inequality in an Online Encyclopedia. Claudia Wagner, David Garcia, Mohsen Jadidi, Markus Strohmaier. ICWSM (2015) </a> </span></div> <br> <div style="font-size:22pt"><span> <a href="https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-016-0066-4"> Women Through the Glass Ceiling: Gender Asymmetries in Wikipedia. Claudia Wagner, Eduardo Graells-Garrido, David Garcia, Filippo Menczer. EPJ Data Science (2016) </a> </span></div> --- # Gender in Wikipedia biographies **We analyzed 1 Million biographies in the top 20 Wikipedia languages** .center[] --- # Words predicting gender in Wikipedia  --- layout: true <div class="my-footer"><span><a href="https://www.pnas.org/doi/10.1073/pnas.1717781115"> Analyzing gender inequality through large-scale Facebook advertising data. D. Garcia, Y. Kassa, A. Cuevas, M. Cebrian, E. Moro, I. Rahwan, R. Cuevas. PNAS (2018) </a></span></div> --- # Gender bias in digital traces: Facebook  **Facebook Gender Divide (FGD) as measured with FB advertisement data** --- # The FGD reflects offline inequality  - Regression model of FGD as a function of WEF gender inequality indices - FGD is a sign of education gender inequality --- # When moderation silences minorities - AI is used automatically in moderation to detect hateful and sexist text - **Do models label non-harmful use of identity and slur terms as harmful?** - Examples of false positives in Hatecheck: - Reported and quoted use: *“I was called a [SLUR] so many times growing up”* - Non-hate expressed using negated hateful statement: *“We shouldn’t let [IDENTITY] suffer”* - Denouncements of hate that quote it: *“Statements like ‘[IDENTITY] are scum’ deeply hurtful.”* - Denouncements of hate that make direct reference to it: *“You have to stop calling [IDENTITY] disgusting.” * **Negative Recall of a model**: the fraction of correctly identified negative cases that use identity terms and slurs in a non-hateful way --- ## Trade offs between performance and neg recall  - ChatGPT has high performance but low neg recall: silencing risk - Perspective has worse performance and neg recall than open source models - Open source models (BERT, FLANT5) have better negative recall <br>.center[**To prevent silencing minorities, we need more than performance**] --- # Summary ### Computational Social Science and AI uses digital traces ### Digital traces can capture gender biases - **Wikipedia biographies of women contain gender stereotypes** - **Facebook data follows gender inequality across countries** ### AI moderation can silence minorities with false positives - **Open source models do better despite lower performance** .center[**More at: [www.dgarcia.eu](https://dgarcia.eu)**] .center[**[Bluesky: @dgarcia.eu](https://bsky.app/profile/dgarcia.bsky.social)**]